The comprehensibility of solutions is critical in AI-supported decision-making processes, particularly in the healthcare sector, where the outcomes of AI techniques must be explained to clinicians to enable informed and reliable decisions. Various approaches have been proposed to address this challenge, primarily by identifying the most influential features in complex models, such as deep neural networks. However, such approaches may oversimplify models, potentially overlooking intricate relationships between clinical, genetic, or other variables. In this paper, we present a rule-based method for deriving diagnostic models from historical patient data. Alongside the rule-generation engine, we introduce a method to assess the impact of each feature on a diagnosis. Additionally, an interactive, intelligent interface generates dashboards that enable users to explore their Explainable AI models and extract meaningful insights. To demonstrate the efectiveness of our approach, we discuss diferent applications, including a use case involving Acute Myeloid Leukemia data.
A
The use of Explainable AI (XAI) in medical data analysis pipelines represents a transformative approach
to enhancing both the accuracy and trustworthiness of healthcare diagnostics and treatment planning
[
XAI methods in healthcare typically focus on generating human-interpretable explanations for AI predictions. These methods include feature importance analysis, rule-based models, and counterfactual explanations. By making AI more interpretable, XAI improves clinical decision-making, increasing trust among healthcare professionals, and enhancing patient safety. Furthermore, regulatory frameworks, such as the EU’s General Data Protection Regulation (GDPR) and the AI Act, underline the need for explainability, making XAI an essential factor in the adoption of AI-supported systems within the healthcare sector.
A crucial aspect of XAI in healthcare is the design of user interfaces that efectively communicate
AI-generated insights to medical practitioners by incorporating visualizations, interactive elements,
and natural language explanations to enhance interpretability [
CEUR
ceur-ws.org
Notable examples include platforms such as NEAR [
In this paper, we propose a machine learning and visualization tool that combines the efectiveness and explanatory power of rule-based systems with the simplicity and immediacy of feature-based visualizations. The proposed approach employs data-driven rules to forecast disease onset in patients, highlighting the complex relations behind the predictions. Moreover, the contribution of each feature to the diagnosis is computed based on the rules. For each patient, a risk score is calculated and decomposed into the contribution of each feature. All this information—the rules, feature relevance, diagnosis, and the contribution of each feature to the diagnosis—is presented in an interactive user interface. The manuscript is structured as follows: Sec. 2 describes the algorithm to generate the rules and determine feature relevance from the data; Sec. 3 presents the tools used to display the results to the end user; Sec. 4 introduces some applications where the approach has been tested; Sec. 5 describes the dashboard developed to visualize the results; and Sec. 6 summarizes the contribution of the presented tools for the XAI community.
Rule-based models are very powerful to explain why a decision was suggested by an AI system. The simplest way to define a rule is by combining a premise and a consequence: IF <premise> THEN <consequence>. The premise contains the conjunction of some conditions about the inputs while the consequence contains the output of the system, i.e. the quantity to be predicted. This could be an example of rule:
IF = {, }
AND < 10 THEN =
As the example shows, conditions could involve both categorical variables (i.e. for which an ordering
cannot be imposed) and numerical attributes. Also, the output could be categorical (in this case the
problem is referred to as classification ) or numerical (regression), even if in this paper we only consider
the first case. Several approaches have been proposed in literature to generate rules from data. The most
popular technique is Decision Tree, that iteratively divides the dataset in smaller subsets. This
divideand-conquer approach provides an easy-to-understand classification, but the classification accuracy is
usually poor. For this reason, ensemble approaches, such as Random Forests, aim at combining several
rule set to achieve a richer and more accurate model of the system. In general, rules could be disjoint
(like in Decision Trees) or overlapping (like in Random Forests). Overlapping rules have been proven
[
Step (c) constitutes the core of the approach and requires proper algorithms to build a Boolean
function that maximizes the accuracy of the model. As a matter of fact, the Boolean function should
avoid overfitting, i.e. excessive adherence of the model to the experimental data, and underfitting, i.e.
poor efectiveness in describing the available data. To reach this goal, the Shadow Clustering algorithm
[
Rules can be characterized by diferent metrics that measure their efectiveness in describing the training set. Each rule is said to cover an example x if its premise matches x. The examples covered by a rule are called positive, while those not covered are called negative. A true positive is a positive example whose output matches the output of the rule, whereas the output of a false positive difers from that of the rule. Similarly, true negative and false negative cases can be defined. The simplest indicators for a rule are covering (or true positive rate) and error (or false positive rate), which account for the fraction of cases correctly and incorrectly described by the rule, respectively. The balance between covering and error is crucial to ensure that rules are general and efective in describing the data. LLM allows for the calibration of the amount of error in a rule, and usually, permitting a small amount of error is beneficial for the quality of the rules.
In addition to these rule-related indicators, metrics for each condition can also be introduced. In particular, we can compute how covering and error change when the studied condition is removed. Since removing a condition corresponds to eliminating a constraint, covering and error usually increase (or remain unchanged). The increase in error, in particular, is a strong indicator of how important the condition is within the rule. If the error increases significantly, the condition cannot be eliminated without heavily afecting the quality of the rule and is therefore essential. Conversely, a small increase in error corresponds to a condition that could be removed without causing significant damage. Let Δ () be the increase of error in rule after removing condition .
The importance of conditions in a rule can, in turn, be used to compute the relevance of attributes associated with each condition. By summing Δ () over all the rules and all the conditions associated with attribute , the absolute relevance ( ) of is computed. If the sum is limited to the rules whose output class is , the relative relevance ( ) is computed.
Additionally, it is possible to compute a relevance ( ) for each subset ⊂ of the domain of by summing Δ () only for the conditions that contains . For categorical inputs this corresponds to evaluate the relevance of each possible value, while for ordered variables, a relevance score can be associated with every interval. Similarly also the relative relevance subset of . ( ) can be associated with each
Understanding and interpreting the rules generated by AI models is crucial, especially in medical applications where explainability is fundamental. The Rulex Platform1 provides multiple tools to visualize, evaluate, and manage rules, making it easier for users to explore the relationships between variables and assess the impact of diferent conditions. In this work we show how four key components can support stakeholders in making more informed decisions: the Feature Ranking, the Rule Viewer, the Rule Manager, and Rulex Studio. These tools are currently used in several fields, such as logistics and ifnancial services, but here we focus specifically on the applications in healthcare problems.
Feature Ranking. The Feature Ranking tool provides a quantitative assessment of how diferent input attributes contribute to the predictive model, directly leveraging the metrics introduced in Sec. 2. Specifically, it utilizes the concept of error variation ( Δ () ) discussed in Sec. 2.2 to determine the importance of each condition within a rule. By aggregating these variations across all rules, the system computes both the absolute relevance of an attribute—reflecting its overall impact on model predictions—and the relative relevance, which quantifies its importance within a specific output class.
The Feature Ranking interface allows users to explore these computed relevance scores interactively. Attributes can be sorted based on diferent criteria, such as their absolute importance or their specific contribution to a given diagnosis. Additionally, for categorical variables, the tool evaluates the relevance of individual values, while for numerical attributes, it highlights key threshold intervals that significantly afect the model’s decision-making process.
Rule Viewer. The Rule Viewer provides an interactive environment for visually exploring rule-based models, ofering intuitive representations of extracted rules and their relationships with input attributes. Figure 1 shows an example of the Rule Viewer applied to the Alzheimer’s Disease dataset presented in Sec. 4.1.
The central rule chart constitutes the core visualization element. The outer circular ring represents input attributes, sorted by their absolute relevance, with each segment corresponding to a specific attribute. Attributes are colored and annotated to distinguish nominal (N), integer (I) and continuous (C) types. Within this outer ring, circles represent individual rules, grouped according to the predicted output class (e.g., Control or Sick). The size of each circle is proportional to the rule’s coverage, while the hole in the center reflects the associated error rate. Hovering over any rule (as shown in the figure for Rule #12) reveals its logical conditions, displays detailed metrics such as covering and error, and highlights the corresponding attributes on the outer ring to visually link the rule to its defining features.
The interactive settings panel on the right-hand side enables users to dynamically adjust visualization parameters, such as the number of attributes displayed, attribute sorting criteria, and relevance thresholds.
Rule Manager. The Rule Manager task ofers a structured interface for inspecting, refining, and optimizing rule-based models. It is especially valuable when working with large rulesets that require manual adjustments. The tool provides a spreadsheet-like environment where users can filter, sort, and modify rules, adjusting conditions or output values as needed. Additionally, a history tracking feature allows easy review and reversal of changes, ensuring flexibility and control over rule modifications.
Rulex Studio. Rulex Studio is an advanced visualization tool that enables users to create interactive dashboards for analyzing AI-generated models. Unlike previous tasks, which focus primarily on rule inspection and refinement, Rulex Studio provides a broader visualization framework, supporting data exploration and model presentation. The main component of Rulex Studio is the View, a customizable workspace where users can design graphical representations of their data. The platform includes various options for adjusting layout configurations, importing views, and integrating plots. This lfexibility makes it an ideal tool for presenting explainable AI models to clinicians and other stakeholders.
Medical AI applications must be designed to meet specific clinical needs, ensuring that predictive models provide transparent, reliable, and actionable insights for healthcare professionals. In this study, we focus on two case studies: Alzheimer’s disease and Acute Myeloid Leukemia (AML), both severe conditions where early detection and prevention are crucial for improving patient outcomes. By applying our rule-based approach to these diseases, we aim to demonstrate how interpretable predictive models can support clinicians in identifying at-risk individuals and making informed decisions to enhance early intervention strategies.
For Alzheimer’s disease, we used a dataset [
To develop an interpretable predictive model, we applied LLM, conducting an extensive fine-tuning of its parameters. We evaluated multiple configurations using a 70-30% training-test split, selecting the optimal setup to balance accuracy and generalization. The final model produced 11 rules for the Control class and 7 for the Sick class, with each rule containing between 1 and 15 conditions. The model’s ifnal performance, summarized in Table 1, demonstrated strong predictive capability, confirming the efectiveness of rule-based learning for Alzheimer’s risk assessment.
Training Test
92.6% 90.6%
93.9% 89.2%
For Acute Myeloid Leukemia (AML), we utilized a dataset derived from the study by the Weizmann
Institute of Science [
This study is part of the broader SInISA project2, which aims to develop predictive screening methods for hematological diseases through genetic and AI-driven analysis. Given the small dataset size and the class imbalance, we conducted fine-tuning of Logic Learning Machine parameters using a 70-30% training-test split, but classification performance was lower compared to the Alzheimer’s model. The ifnal model generated 40 rules, with 24 rules for the AML class and 16 for the Control class, each containing up to 10 conditions. Despite the classification challenges, the extracted rules aligned with ifndings from the Weizmann study, reinforcing the known biological patterns: mutations in certain genes (e.g., DNMT3A, TET2, SRSF2, ASXL1, TP53) are commonly found in healthy aging individuals but exceed a critical threshold in those likely to develop AML. This result supports the SInISA project’s goal of identifying early genetic markers for leukemia risk, highlighting the potential of rule-based models in predictive medicine.
To make the results of explainable models more accessible and clinically meaningful, we developed an interactive dashboard using Rulex Studio, a flexible environment for building custom data visualizations. The dashboard supports both model evaluation and in-depth exploration of individual predictions, with the primary goal of providing transparent, patient-specific risk assessments.
A key innovation of our approach is the transition from binary classification to the computation of a personalized disease risk. Based on the analysis of the Logic Learning Machine rule conditions and feature relevance, each patient is assigned a percentage risk score that quantifies the likelihood of disease onset. The following subsections illustrate the dashboard’s main functionalities, using examples based on the Alzheimer’s use case.
The central feature of the dashboard is the ability to compute and visualize a personalized disease risk for each patient. The patient’s risk is calculated starting from a baseline risk, derived from the training dataset. Each attribute of the patient contributes positively or negatively to this baseline, based on the relevance of its specific value. These contributions are derived from the relevance analysis of Logic Learning Machine at the condition level.
As illustrated in Fig. 2, the top-left panel shows the final risk score, with a visual indicator of how
far it deviates from the base value. Below, a radial contribution chart displays the impact of each
attribute, distinguishing between those increasing (red) and decreasing (green) the predicted risk. On
the right, a dual plot shows the cumulative efect of each variable, sorted by relevance: the upper
section tracks the progressive build-up of the final score, while the lower plot quantifies the individual
contributions of all features. This representation provides a clear explanation of both the prediction and
its underlying rationale. A similar strategy has been adopted in platforms such as NEAR [
Beyond evaluating existing patients, the dashboard allows users to manually edit or create new patient profiles, enabling powerful “what-if” analysis. This feature is especially useful for exploring hypothetical situations and assessing how changes in specific attributes influence the final risk of the disease.
Each time a variable is modified, the system instantly recalculates the overall risk and updates the visual explanations, allowing users to quickly assess the impact of individual changes. This makes it possible to simulate the efects of lifestyle or clinical interventions and to identify which variables drive risk the most, supporting more informed and personalized prevention strategies.
The dashboard also includes a report generation feature that produces a two-page summary for each patient. This report consolidates all key insights: patient’s input data, predicted risk score, the contribution of individual features, and supporting visualizations such as risk indicator and feature contribution chart. Each section of the report is accompanied by clear captions explaining how to interpret the data, making it suitable for sharing with clinicians or for integration into broader diagnostic workflows. Reports can be generated for patients in the dataset or for manually created profiles, supporting both retrospective analysis and scenario testing.
While the main focus of the dashboard is on personalized patient analysis, it also includes a global view summarizing the behavior and internal logic of the model. This view provides access to essential evaluation metrics, such as feature ranking and rule visualization, which ofer insight into how the model generalizes across the dataset. The feature ranking panel highlights the attributes that most influence the model’s predictions, computed using the relevance scores assigned by the Logic Learning Machine. These scores reflect the importance of each feature value in shaping the classification rules and, consequently, the calculated risk. The ranking can be customized and sorted according to diferent criteria. The rule manager displays all the rules extracted by the model, color-coded by class and organized by complexity and coverage. Users can filter rules based on the attributes they include, inspect their logic, and visualize associated metrics such as covering and error.
This work presents tools to enhance transparency in AI-supported medical diagnosis. The rule-based approach is essential to convey complex relations in a human-readable format. Nonetheless, efective AI algorithms may not be suficient to enhance trust among decision-makers. The approach proposed in this work aims to overcome this limitation by providing both data-driven rules to support the diagnosis and a user interface to enable stakeholders to utilize the derived insights. The combination of powerful XAI algorithms and an interactive interface may ultimately foster users’ trust, as shown in the two case studies presented. During the preparation of this work, the authors used GPT-4 in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.